OpenAI service tiers, explained
batch, flex, standard, fast,
ultrafast — five tiers spanning a 12× spread in price,
and a decision most teams never make at all.
Short version: the tier is a per-request parameter. Pick the slowest tier your workload tolerates — Batch and Flex are half price, Fast is double, Ultrafast is six times. Standard is the correct default only for interactive traffic you cannot predict.
The five tiers
| Tier | service_tier | Rate | Speed | Residency |
|---|---|---|---|---|
| Batch | batch | 0.5× | ≤24 h | EU on Luna; not stated for Sol, Astra |
| Flex | flex | 0.5× | queued | EU on Luna; not stated for Sol, Astra |
| Standard | default | 1× | 1× | EU + global |
| Fast | fast / priority | 2× | up to 2.5× | EU on Sol and Luna; not on Astra |
| Ultrafast | ultrafast | 6× | fastest, no published multiple | Sol: US, EU, global. Astra: US, global |
Rate is the published multiplier on the Standard per-token price. Speed is the vendor's published ceiling where one exists, not a guarantee for your workload. Residency is per model, from OpenAI's Fast and Ultrafast guides and model pages, checked 27 October 2026.
What the multiplier actually does to a bill
The multiplier applies to all four lines of the usage dashboard — uncached input, cache
reads, cache writes and output — so it scales the whole request, not one part of it. On
GPT-6 Astra, whose standard rates are $10 input and $50 output per
million tokens, Ultrafast publishes $60 input and $300 output —
six times on every figure.
That last number is worth sitting with. Astra on Ultrafast costs $300 per million output tokens. GPT-6.1 Sol on Standard costs $10. The same workload can be thirty times more expensive purely because of a parameter nobody on the team remembers setting.
Where Fast and Ultrafast can run depends on the model
Residency is not a property of the tier. On GPT-6.1 Sol, both Fast and Ultrafast support EU data residency. On GPT-6 Astra, neither does: Fast is not offered with EU data residency, and Astra Ultrafast runs on US data residency and global processing only. GPT-6 Luna supports EU residency on Standard, Fast, Flex and Batch, and has no Ultrafast tier.
If your contracts specify where inference happens, the model decides whether a premium tier is on the table. An EU-bound workload that needs speed has an answer on Sol and none on Astra. Earlier versions of this page said neither premium tier had an EU endpoint; that was true of Astra only, and has been corrected.
Ultrafast also starts throttled
Ultrafast has its own rate limits, separate from Standard and Fast — unlike Fast, which draws on the same limits as Standard. OpenAI's default token-per-minute limits for Ultrafast are 500K, 1M and 5M for GPT-6 Astra across its Build, Launch and Grow usage tiers, and 1M, 4M and 40M for GPT-6.1 Sol. A large job can be throttled on Ultrafast even when your Standard limits are generous, so check the limits page before you size the run.
Picking a tier, by workload shape
- Nobody is waiting. Nightly enrichment, backfills, eval runs,
embeddings refresh →
batch. Half price, and the largest safe saving available to most teams. - Traffic is spiky and retryable. Non-production endpoints, batch-ish
interactive work →
flex. Also half price, but it may queue. - A human is waiting, volume unpredictable. →
standard. Cache the system prompt instead of buying speed. - A human is waiting and the latency is the product. →
fast, set per request. - Wall-clock time is the constraint and volume is large. →
ultrafast, on GPT-6.1 Sol or GPT-6 Astra — Sol first if you need EU residency or a lower bill.
A five-day order of operations
Tier changes are cheap and reversible, which makes them the right first move — before any prompt engineering or model swap:
- Sort every workload by one question: is a human waiting for this response?
- Move the "no" pile to
batch. Watch the async failure rate. - Move the "maybe" pile to
flexbehind a feature flag, and watch retries. - Leave the "yes" pile on
standard, and cache its stable prefix. - Only then look at
fast— for the specific requests where latency is actually worth 2×, not as an account-wide default.
Work out your own number
The composer on the home page takes your token counts, cache hit rate and daily volume, and shows the per-call and monthly cost on every tier at once. It runs entirely in your browser — nothing is uploaded.